The authors [PS49] show that for both long term forecasting for planning and short term forecasting for online power system operation, simpler speedy and robust forecasting software is very useful. For this purpose, an ANN plus fuzzy logic is shown to be very appealing.
Based on the new developed structure variable ANNs, two models-the daily peak load (DPL) model and daily 24-hour load (DHL) model are proposed in the paper [PS50]. The cluster Gaussian analysis is used for the training of the models. The effectiveness of the new forecasting strategy is demonstrated by training and testing using the data collected from the Jing-Jin-Tang network.
The paper [PS51] presents an approach to power load forecasting using ANN. Based on weather conditions and past history of load consumption, a load forecast is made by the utility companies to deliver the appropriate load to its customers. Power systems operation and planning functions such as unit commitment, security analysis, state estimation, etc. are benefited with an accurate load forecast. Improving the accuracy of the load forecast can save a significant amount of money. ANNs permit adaptability to climate changes compared to other forecasting methods in use. The results obtained by using ANN have been found to give better results than other conventional techniques.
One of the most widely established forecasting methods is based on the ARMAX processes (Auto-Regressive-Moving-Average with eXogenous variables). This method is mathematically well known and widely used in practice. In the paper [PS52] the authors present a type of forecasting model based on a multilayer perceptron which contains approximating ARMAX models. They show the advantages it can bring and describe an algorithm which aims at reducing the number of parameters associated with such a model. A criteria for comparison is also proposed, taking into account both the performance of the model in terms of the adjustment (residual variance value) and the number of parameters used (number of synaptic weights) for a given number of observations. After having carried out several simulations, the authors have applied these results and modelled the daily electrical consumption series in metropolitan France. The neural model obtained is more powerful in terms of adjustment for a reduced number of parameters than the corresponding linear model.
The paper [PS53] investigates the effectiveness of the ANN approach to short term load forecasting in electrical power systems. Using examples, the learning process and capabilities of an ANN in the prediction of peak load of the day are demonstrated. Different data normalizing approaches and input patterns are employed to exploit the correlation between historical load and temperatures and expected load patterns. A number of ANNs are included with emphasis given to their practical implementation for electrical power system control and planning purposes. The networks have been trained on actual power utility load data using a backpropagation algorithm. The prospects for applying a combined solution using ANNs and expert systems, called the expert network are also discussed. Consideration is given to expert networks as a more complete solution to the forecasting problem which neither system alone can provide.
An adaptive ANN based short-time electric load forecasting system is presented [PS54]. The system is developed and implemented for Florida Power and Light Company (FPL). Practical experiences with system are discused. The system accounts for seasonal and daily charackeristics, as well as abnormal conditions such as cold fronts, heat waves, holidays and other conditions. It is capable of forecasting load with a lead time of one hour to seven days. The adaptive mechanism is used to train the ANNs when on-line. The overall system is developed in two steps. The first step is to produce an adequate response of the ANN up to a certain year using past data. The second step is to integrate the ANN weights obtained in the first step into an adaptive mechanism. This adaptive procedure along with other periphereal modules serves as the nucleus of the real time execution mode of the overall system. Data preparation, selection of appropriate ANN architecture and input variables, and training of a multi-layered perceptron type ANN using back-propagation algorithm, fall under the first step. Hourly load and weather data for years 1984 to 1989 was provided by Florida Power and Light Company. The data was first diveded into four categories based upon seasonal characteristics, giving winter, summer, transition I (March) and transition II (October), then it was further categorized based upon the type of the day of week such as weekend, weekady or Monday. Similairly each day was divided into five periods to account for different level of the commercial/industrial activities that might be present. Some days were handled specially, those having extreme temperatures such warm and cold fronts, and holidays. Thus a total of 212 different ANN architectures were required for each category of the data, 80 as to cover every possible conditions. As the demand for load increases from year to year, this growth in demand is calculated that takes into account the temperatures of the years under consideration, thus immunizing the growth-ratio from temperature variations between the years. The ANN is trained using weather and load variables derived from the above data to determine an optimal architecture and set of variables for each possible condition. At this stage an adaptive mechanism can be incorporated into the ANN, and it was integrated into a user interface that automatically selects the appropriate architecture to use for making the forecast, adapts for new actual load and temperature data and does post-processing on the forecast load. The adaptive mechanism presented show a method by which new data can be incorporated into the old training set without adverse effect of accuracy. Three different types of adaptation were used: Daily, monthly and weekly adaptations. Daily adaptation uses the weights obtained from the conventional network as the initial weights. Each day is adapted by a few training iterations with the new data in which weights are changed subject to certain constraints. Daily adaptation improves the weights for a short term only, after some time it becomes counter-productive due to data memorization problem. Therefore at the end of each month, the initial weights are re-adapted with the whole set of new data for the new month in a single training session. This proces is done in a matter of seconds. Monday and weekends are adapted on veekly basis. Accurate forecasting for cold fronts, warm fronts and holidays is of special importance to utility companies. In the load forecasting system presented here, temperature ranges below 55F were categorized as cold fronts and temperature ranges above 90F were treated as warm fronts. Forecasting for such conditions was handled by devising ANN architectures and training sets for each of these conditions. Furthermore, holiday data from the past 5 years together with data for the weekends were used to formulate the training data set. For the weekends, data were taken from the previous two weeks before a particular type of holiday. By collecting similar days from the past, the authors ensure that the characteristic of only that type of holiday is reflected in the data set. Before making a forecast, the weights of the ANN are fine-tuned to the characteristics of the forecast day's conditions. This process does not affect the weights on a permanent basis because they are discarded after the forecast is made. 2-day to 7-day ahead forecast were also developed for the system presented there. Various ANN architectures were designed and trained for this case. The number of input variables varies from 5 to 16, depending on the season, period and type of day being forecasted. The architecture for a particular period of a certain day type is constant for the whole sesson. There are 15 architectures for each season in each day ahead forecast. Each season is divided into three day types. Each day type has 5 architectures for each period. The overall system consists of two modules. One is the system forecast initialization (SFINIT) module and the other is the real time execution moduel. The SFINIT module is used to train various ANN architectures and obtain weights for different years, seasons and months. This process is done off-line. The real time mode utilizes the ANN weights obtained using the SFINIT module and allows the user to interactively obtain a forecast on-line. Forecast for 1 hour to 7 days can be obtained based on forecast temperature. The system allows the user to make adjustments in the forecast. This facility gives the load forecaster flexibility to adjust sudden changes in weather conditions. The system operation starts with the selection of period to train and ANN architecture. The next step is the training of the ANN followed by the testing. Once the ANN is trained, the system is placed on-line for forecasting. Adaptation is done before making the load forecast. The selection of the period can be for a whole year, a particuler season, a month, or a particular period. This selection is carried out for 1 to 7 days architectures.
Send Questions, Comments, and Suggestions to bencr@hotmail.com